EU AI Act — Art. 4 & Art. 50
AI transparency
Effective 2026-07-25. Art. 50 transparency obligations apply from 2026-08-02.
Altvisor generates text with AI. Under Regulation (EU) 2024/1689 (Artificial Intelligence Act) we tell you where, which kind of model, where it runs, and how a person stays in control of what it produces.
This page describes how Altvisor itself uses AI. It is a disclosure, not legal advice, and it does not tell you what your own business must do.
The roles we hold
AI Act obligations attach to roles, not to "anyone using AI". We are a provider of the systems described below, because we put them on the market under our own name, and a deployer of a third-party model, which we run under our own authority. We are not a provider of a general-purpose AI model.
The AI systems we operate
Every system below exists in the product today. Removing a system removes its row.
| System | What it generates | Model class and hosting | Human review | Articles engaged | Disclosed in the interaction |
|---|---|---|---|---|---|
| Alt-text generationThe authenticated app (upload, bulk import, candidate review), the integration API, and the e-commerce plugins. | Synthetic text describing an image, in the requested EU language, plus a decorative/non-decorative determination. | EU-hosted multimodal large language model (Mistral, La Plateforme)France (EU) — Mistral La Plateforme. No inference leaves the EU on any plan. | Every generated alt-text is editable and versioned: an edit inserts a new row and supersedes the old one, so the text active on any date is reproducible. The review console records who approved what and when, and that record is exportable as an AI provenance report. | Art. 50(2), Art. 4 | Not applicable — no direct interaction |
| Public alt-text demoThe anonymous demo widget on the marketing site. | Synthetic alt-text for an image the visitor uploads, shown immediately in the page. | EU-hosted multimodal large language model (Mistral, La Plateforme)France (EU) — Mistral La Plateforme. | The demo produces a suggestion only. Nothing is published; the visitor decides whether to use the text. The uploaded image is processed in memory and never retained. | Art. 50(1), Art. 50(2), Art. 50(5), Art. 4 | Yes |
| Compliance-report prose enrichmentScan reports and EAA Annex V accessibility statements generated for a customer. | Optional prose that expands on a deterministic, in-repo remediation baseline. Never the conformance grade, the legal citations, the fine figures, or the compliance status — those are templated and deterministic. | EU-hosted large language model (Mistral, La Plateforme)France (EU) — Mistral La Plateforme. | Model output is admitted only when it passes a plain-prose guard; anything flagged, empty, structured, or over-long is dropped and the deterministic baseline stands. Draft statements are edited and published by a human, who holds editorial responsibility for the published text. | Art. 50(2), Art. 4 | Not applicable — no direct interaction |
The model provider is disclosed as a subprocessor. Mistral AI (La Plateforme)
What we do not operate
- We do not run an AI chatbot. Our support chat routes to a human. If that ever changes, the AI nature will be disclosed inside the chat window itself — not in these terms.
- We do not generate or publish AI-generated imagery, so no Art. 50(4) deepfake disclosure arises on our own surfaces.
- We do not operate emotion-recognition or biometric-categorisation systems, so Art. 50(3) does not apply to us.
What Art. 50 does and does not require
The market is currently being told that all AI-generated content must be marked. That is broader than the law. Because we sell precision about this regulation, we state the boundary plainly — including where it works against us.
| Claim | What the article says |
|---|---|
| A chatbot must disclose that it is AI. | Correct (Art. 50(1)). The disclosure must be perceivable in the interaction itself — text buried in terms and conditions, or a vague label like "assistant", does not satisfy it. |
| All AI-generated content must be marked. | Too broad. The machine-readable marking duty in Art. 50(2) binds the provider of the generative system — the model vendor — not every business that uses one. |
| AI-generated product images must be disclosed. | Arguably in scope. Art. 50(4) covers photorealistic AI imagery that could pass as authentic, and commercial product photography gets no artistic exemption. Upscaling, colour correction, and relighting are assistive editing and stay outside it. |
| AI-written product descriptions and alt text must be disclosed. | Not in scope. Art. 50(4) second paragraph covers text published to inform the public on matters of public interest, and expressly exempts content that has undergone human review with someone holding editorial responsibility. |
| Only personal use is exempt. | The wrong frame. Obligations attach to roles (provider, deployer) and to system types — not to a personal-versus-commercial distinction. |
| Fines reach €15M or 3% of turnover. | Correct (Art. 99(4)). For SMEs and start-ups the cap is the lower of the two figures, not the higher (Art. 99(6)). |
Why an accessible disclosure is not optional
Art. 50(5) requires transparency information to be provided at the first interaction and to conform to the applicable accessibility requirements. An AI disclaimer a screen reader cannot announce fails the AI Act and the European Accessibility Act at the same time. Every disclosure surface in this product is built to WCAG 2.1 AA for that reason.
AI literacy (Art. 4)
Art. 4 has applied since 2025-02-02 and asks for measures proportionate to the operator's size and context. No certification exists or is required. This is the record of ours.
- 2026-07-25 — Mapped every AI system in the product to a role (provider / deployer) and the articles genuinely engaged, and recorded the mapping in this file so it is reviewed whenever a system changes.
- 2026-07-25 — Documented and enforced the boundary the model may not cross: no model authors a conformance grade, a legal citation, a fine figure, or a compliance status. Those are deterministic and in-repo.
- 2026-07-25 — Verified that every AI disclosure surface meets WCAG 2.1 AA, as Art. 50(5) requires transparency information to conform to the applicable accessibility requirements.
- 2026-07-25 — Briefed on the difference between the obligations of a generative-model provider and those of a business deploying one, so neither our product copy nor our support answers overstate a customer's duties.
Human oversight and how to reach us
Every AI output in this product is a suggestion until a person accepts it, and the record of who accepted what is retained and exportable. To question an AI-generated output, or to ask for it to be reviewed by a person, write to hi@altvisor.eu.